docs: add ainetcafe (Kimi K3) provider page

docs.json

@@ -112,6 +112,7 @@
                     "group": "LLM Providers",
                     "pages": [
                       "openhands/usage/llms/openhands-llms",
+                      "openhands/usage/llms/ainetcafe",
                       "openhands/usage/llms/aws-bedrock",
                       "openhands/usage/llms/azure-llms",
                       "openhands/usage/llms/google-llms",

openhands/usage/llms/ainetcafe.mdx

@@ -0,0 +1,54 @@
+---
+title: Use ainetcafe (Kimi K3) with OpenHands
+description: Configure the ainetcafe OpenAI-compatible endpoint for Kimi K3 in Agent Canvas or OpenHands.
+---
+
+Use ainetcafe when you want Kimi K3 through an OpenAI-compatible endpoint. [ainetcafe](https://ainetcafe.com/k3/) serves Kimi K3 from its own cluster at the released MXFP4 precision, with a 256K context by default (1M on request), tool calling, image input and prompt caching. OpenHands talks to it through LiteLLM's OpenAI-compatible provider, so no extra dependency is needed.
+
+## Get an API key
+
+1. Create an account at [ainetcafe](https://microquickjs.com/register?lng=en). New accounts receive a small sign-up credit.
+2. Open `Token Management`, create a token and copy the key (it starts with `sk-`).
+
+## Configure Agent Canvas
+
+1. Open `Settings > LLM`.
+2. Switch to the `Advanced` tab.
+3. Set the following:
+
+| Setting | Value |
+|---|---|
+| Custom Model | `openai/Kimi-K3` |
+| Base URL | `https://microquickjs.com/v1` |
+| API Key | Your ainetcafe key |
+
+4. Save the profile, start a new conversation and send a short message to verify the model responds.
+
+The model ID is case-sensitive: use exactly `Kimi-K3`.
+
+## Configure with environment variables
+
+For headless or SDK runs, the same settings map to LiteLLM's environment variables:
+
+```bash
+export LLM_MODEL=openai/Kimi-K3
+export LLM_BASE_URL=https://microquickjs.com/v1
+export LLM_API_KEY=sk-your-ainetcafe-key
+```
+
+## Notes
+
+- K3 reasons before it answers. The first token on a large repository can take a few seconds; later tokens stream normally. Reasoning is returned in `reasoning_content` and does not count against your visible output.
+- Prompt caching is on by default. Repeated context in long agent sessions is billed at the cached-input rate.
+- Current pricing, a live availability probe and a comparison with other K3 providers are published at [ainetcafe.com/k3](https://ainetcafe.com/k3/).
+
+## Fix Common Problems
+
+| Problem | Check | Next step |
+|---|---|---|
+| `404 model not found` | The model ID | Use exactly `openai/Kimi-K3` in `Custom Model`. |
+| `401 Invalid token` | The API key | Re-copy the token from `Token Management`; keys from other sites do not work here. |
+| `403 insufficient balance` | Wallet balance | Top up in `Wallet`, or wait for a plan quota reset. |
+| Slow first token | Reasoning depth | Expected on large contexts; keep `API_TIMEOUT` generous (10 minutes is safe). |
+
+For the broader configuration decision, see [Configure a Model](/openhands/usage/agent-canvas/model-configuration). For LiteLLM's provider behavior, see [LiteLLM's OpenAI-compatible documentation](https://docs.litellm.ai/docs/providers/openai_compatible).